eval: emit diagnostic failure_reason + JSONL feedback export from all evaluators

[wildcard]

Context

Learning L2 from docs/research/dspy-prompt-optimization.md (PR #1448): DSPy's GEPA optimizer needs a metric that returns a score and textual feedback — the feedback is what a reflection model reasons over when proposing a better prompt. It is also simply better failure reporting for humans.

Today the CI eval target (cargo test --test evaluation, wired to src/evaluation/) scores per-rule boolean pass/fail. EvaluationResult.failure_reason (src/evaluation/models.rs) exists and the rule evaluators populate it, but the strings are prose for a report, have no stable shape, and nothing exports them per case.

Problem

• src/evaluation/evaluators/correctness.rs → evaluate_exact_match / evaluate_command_equivalence / evaluate_pattern_match return (bool, Option<String>) with free-text reasons; posix.rs and safety.rs follow the same pattern.
• There is no machine-readable per-case record of (input, expected, actual, score, feedback), so neither an optimizer nor a triage script can consume eval failures. Aggregate pass rate is the only artifact.

Proposal

1. Diagnostic feedback with a stable shape from every evaluator: which rule failed, expected vs actual, the flag/operand diff, and for safety cases the pattern id that fired (or should have).
2. JSONL export per run — one line per case: {test_id, backend, category, input, expected, actual, passed, score, feedback}. Follow the src/evaluation/sft_export.rs pattern (pure transform from EvaluationResult + TestCase, no IO inside the module; the harness CLI writes the file via a --feedback-jsonl <path> flag or alongside --format json).
3. No change to pass/fail semantics — this is additive; pass rates on main must not move.

Acceptance criteria

• Every failing EvaluationResult carries a failure_reason naming the failing rule and an expected/actual pair.
• cargo test --test evaluation -- --backend static_matcher --feedback-jsonl out.jsonl writes one valid JSONL line per evaluated case.
• Regression guard: unit tests in src/evaluation/evaluators/ asserting the feedback shape for each rule, plus a contract test that JSONL lines round-trip through serde.
• Pass rate for static_matcher unchanged.

Related

• docs: add DSPy prompt-optimization learnings and adoption roadmap #1448 — the research doc (see L2, and §2.3 for the metric-attribution detail: mirror src/evaluation/evaluators/, not the unlinked ladder in tests/evaluation/src/evaluator.rs)
• Improve embedded model command generation quality (prompt engineering + fallback) #798 — open umbrella issue for embedded prompt quality
• WP10-11: Prompt Engineering Framework (Versioning & A/B Testing) #517 — closed predecessor (WP10-11 Prompt Engineering Framework)
• Prerequisite for the Phase 2 DSPy harness (tools/dspy-harness/) proposed in the doc.